§ 01 — Reader's reference

A glossary of the terms in question.

Industry-standard definitions for the technical and governance vocabulary Marici uses, set alongside how Marici itself uses each term. This page does not assert misuse — it gives a reader the means to compare and judge for themselves.

Artificial Intelligence (AI)

Industry-standard usage

A set of computational techniques — most prominently, machine learning and large language models — for producing outputs (classifications, predictions, generated text or images) from data. In industry practice the term is meaningful only when paired with a description of the system: what model architecture, trained on what data, evaluated against what benchmark, and operated by whom.

OECD AI Principles; NIST AI Risk Management Framework (AI RMF 1.0, 2023)

How Marici uses the term

Marici uses "AI" as a stand-alone superlative — "the world's most sophisticated AI," "50 AI tools," "AI Intelligence Analyst," "AI Prosecutor," "AI Behavioral Scientist" — without naming a model architecture, vendor, training dataset, evaluation benchmark, or engineering team for any of the tools so described.

Source on marici.org [archived]

Large Language Model (LLM)

Industry-standard usage

A neural network — typically a transformer with billions or trillions of parameters — trained on very large text corpora to produce natural-language outputs. Major LLMs (Claude, GPT, Gemini, Llama) are released with model cards specifying architecture, training-data scope, evaluation benchmarks, and intended use.

Anthropic Claude model card; OpenAI GPT-4 technical report; Google DeepMind Gemini technical report; Meta Llama model card.

How Marici uses the term

Not used as a term by Marici. The publication does not name any LLM provider, model family, or model version anywhere in its public material reviewed.

Source on marici.org [archived]

Predictive mapping / predictive policing

Industry-standard usage

The use of statistical or machine-learning models to forecast where or against whom a crime is likely to occur. Accuracy is a contested term in this literature: the academic record documents systematic bias, feedback loops in self-reinforcing data, and high false-positive rates relative to base trafficking rates. "80%+ accuracy" is a number whose meaning requires the underlying confusion matrix — true positives, false positives, precision and recall — and a published baseline rate.

Lum & Isaac (2016) "To predict and serve?"; AI Now Institute predictive-policing reports; NIST AI RMF.

How Marici uses the term

Marici claims "predictive mapping at 80%+ accuracy" without disclosing what is being predicted, on what data, against what baseline rate, or how accuracy is measured.

Source on marici.org [archived]

Behavioral science

Industry-standard usage

The empirical study of human behaviour, drawing on psychology, neuroscience, economics, and sociology. In an applied context — counter-trafficking, public health, security — practitioners publish methodologies, IRB approvals, and outcomes against pre-registered hypotheses, and identify their principal investigators.

American Psychological Association methodological standards; APA / NIH IRB guidance.

How Marici uses the term

Marici references "behavioral science" both in its homepage AI claim ("merging the world's most sophisticated AI with criminology and behavioral science") and in the named tool "AI Behavioral Scientist." No principal investigators, methodology, IRB approvals, or published studies are identified for these claims.

Source on marici.org [archived]

Full-stack

Industry-standard usage

Originally an engineering term for a developer or system that covers every layer of an application — front end, back end, database, infrastructure. Extended in AI contexts to mean an organisation that owns the entire AI lifecycle: data collection, model training, evaluation, deployment, monitoring, and downstream product. "Full-stack AI" claims by laboratories (e.g. OpenAI, Anthropic, Google DeepMind) are paired with disclosure of each layer's vendors, staffing, and compute footprint.

Industry usage across vendor whitepapers and academic AI infrastructure literature.

How Marici uses the term

Marici describes itself as building "the world's first 'full-stack' AI nonprofit." No public disclosure could be located of the model providers, infrastructure vendors, dataset sources, evaluation pipelines, monitoring systems, or engineering staff that would constitute the layers of any full stack.

Source on marici.org [archived]

Benchmark

Industry-standard usage

A standardised test set used to compare the quality of AI systems against one another. Public benchmarks for general-purpose models include MMLU, GPQA, HumanEval, GSM8K, BIG-bench, and HELM; domain-specific benchmarks exist for vision, robotics, biology, code, and so on. A claim of "state of the art" or "the most sophisticated" is, in industry usage, falsifiable against benchmark scores.

Hendrycks et al., MMLU; Rein et al., GPQA; Chen et al., HumanEval; Stanford CRFM HELM.

How Marici uses the term

Marici claims "the world's most sophisticated AI" and "80%+ accuracy." No public benchmark results, evaluation methodology, or comparative scores against any standard test could be located.

Source on marici.org [archived]

Whitepaper / technical report

Industry-standard usage

A technical document describing the architecture, training, evaluation, and intended use of an AI system. Major laboratories publish whitepapers for every frontier model; specialised AI nonprofits — including counter-trafficking peers such as Thorn — publish technical case studies of their tools.

Thorn's published Spotlight engineering case studies; OpenAI / Anthropic / Google DeepMind model reports.

How Marici uses the term

No whitepaper, technical report, or engineering case study could be located in Marici's public material or any third-party academic / industry repository.

Source on marici.org [archived]

Peer review

Industry-standard usage

The process by which a scientific claim is examined by independent experts in the field before it is published in a journal or conference proceedings. Peer review is the gating mechanism that lets readers distinguish a substantiated finding from a press release. For AI, primary venues include NeurIPS, ICML, ICLR, ACL, FAccT, and discipline journals.

NeurIPS, ICML, ICLR, ACL, ACM FAccT; PubMed-indexed journals for applied health/behavioural sciences.

How Marici uses the term

No peer-reviewed paper authored by or evaluating Marici's stated AI systems or its claimed Stanford partnership could be located on Google Scholar, ResearchGate, arXiv, or any major conference proceedings.

Source on marici.org [archived]

Model card

Industry-standard usage

A standardised disclosure document published alongside an AI model, originally proposed by Mitchell et al. (2019). A model card states the model's intended use, training data scope, evaluation results, known limitations, and ethical considerations. Model cards are now an industry norm for any model offered to third parties, and are required for federally regulated deployments under several emerging AI-governance regimes.

Mitchell et al., "Model Cards for Model Reporting" (FAT* 2019); Hugging Face model-card schema; NIST AI RMF generative AI profile.

How Marici uses the term

No model card, intended-use statement, dataset disclosure, limitations or risks document could be located for any of the named tools ("AI Intelligence Analyst," "AI Prosecutor," "AI Behavioral Scientist") or the unnamed remaining ~47 tools.

Source on marici.org [archived]

Evaluation / accuracy

Industry-standard usage

The empirical measurement of an AI system against held-out test data, an external benchmark, or a real-world deployment. In rigorous AI practice, an accuracy number is meaningless without (a) the evaluation set, (b) the metric definition (precision, recall, F1, calibration), (c) the baseline against which improvement is claimed, and (d) confidence intervals or significance tests.

Liang et al., Stanford CRFM HELM evaluation framework; standard ML literature on confusion matrices and ROC/PR curves.

How Marici uses the term

Marici states "predictive mapping at 80%+ accuracy" without disclosing the evaluation set, the metric, the baseline, confidence intervals, or whether the figure is an internal claim, a third-party measurement, or a projection.

Source on marici.org [archived]